It’s the difference between “asking the LLM” and “using the LLM” (as an agent). If you prompt an LLM with a question it’ll try to answer it using it’s own training (which could contain hallucinations). If you ask an LLM to search the Internet on your behalf, that’s actually quite useful.